Nebius's Run Rate Says $3 Billion. Its Income Statement Says $582 Million and a Loss.

Nebius ($NBIS) reported the second quarter of 2026 before the open on August 12. Group revenue was $582.3 million, up 454% year over year and 46% sequentially. Annualized run-rate revenue reached $3.0 billion. Adjusted EBITDA was $236.2 million, a 41% margin, against a $21.0 million loss a year earlier.

The company also spent $5.66 billion on property and equipment in the same three months, and reported a $190.4 million net loss.

Every one of those is a real number from the same filing. Which one you lead with is the entire argument about this company.

Before the numbers: the history changed

One thing has to be said before any growth rate is quoted. Following the Q2 2025 Toloka deconsolidation, Nebius reclassified prior periods to discontinued operations, and the comparatives in this release do not match what was previously reported:

Our own series carried the old basis until this print; it now carries the filed one. Any year-over-year growth rate for Nebius published before the reclassification was computed against a different company. The 454% headline is correct on the restated base — against the old $152.8M it would have been 281%.

That is not a small distinction for a business whose valuation rests almost entirely on the growth rate.

The quarter

Operating leverage is genuinely arriving. Every expense line fell as a share of revenue — cost of revenues 29% → 23%, product development 41% → 33%, SG&A 65% → 30%, D&A 72% → 45%. Total operating costs went from 206% of revenue to 130%.

And the core business is better than the group. Nebius AI cloud revenue was $574.9M — about 98% of the total — with segment adjusted EBITDA of $285.7M at a 49.7% margin. The cloud alone earned more adjusted EBITDA than the consolidated company, which means Avride and TripleTen cost roughly $50M of EBITDA in the quarter.

Two things sit between that 49.7% margin and the GAAP loss: $119.1M of interest expense (from $4.8M) and $102.5M of share-based compensation, of which $115.9M in product development was non-recurring expense tied to the Eigen AI acquisition.

There is also a quieter accounting change worth naming. Nebius raised the useful life of server and network equipment from four years to five, starting in 2026. D&A is the largest cost line in this business and fell from 72% to 45% of revenue; part of that improvement is the estimate, not the operations.

The capex

Capex was 9.7x revenue. Operating cash flow was strongly positive — $2.25B, against negative $167.7M a year ago — and free cash flow was still negative $3.4 billion.

Against Nebius's Rule of 40: 454.0 of revenue growth plus a free-cash-flow margin of −585.8% gives a score of roughly −132. That is the honest arithmetic and it is also the point at which the ratio stops being informative. A company converting equity and prepayments into GPUs at ten times its revenue is not a Rule of 40 company yet; it is a construction project with a revenue line attached.

How it is being funded — the actual story

The capex is being paid for in a way that did not exist a year ago:

• ~70% of deals closed in Q2 included customer prepayments — an all-time high — covering 50–60% of the associated capex.
• Deferred revenue reached $5,975.2M ($979.4M current, $4,995.8M non-current), from $1,577.5M at the end of 2025. That $4.4B build is the prepayment model showing up on the balance sheet.
• Management expects over $9 billion of customer prepayments in 2026 and cites more than $40 billion of customer commitments.
• An ATM program was used for the first time: 12.7 million Class A shares at a weighted-average $223.60, roughly $2.8 billion gross, with 12.3 million shares still available.
• In July, a first secured debt financing of $775M at SOFR + 2.50%, backed by deployed GPU infrastructure and contracted cash flows from an investment-grade customer.

Total debt is now $8,545.7M against $10,340.5M of shareholders' equity, on a balance sheet that grew from $12.4B to $28.0B in six months.

The number that changed the economics

Buried in the shareholder letter is the figure that matters more than the growth rate:

Together, the expected payback period for the associated capex and related operating costs for Q2 deals is 1 year and 10 months, down from our two-to-three year payback period previously.

A GPU fleet depreciated over five years that pays back in twenty-two months is a different asset from one that pays back in thirty-six. The other deal terms point the same way:

• ACV above $20M per MW on Q2 deals, against a $12M 2026 base, with more than 30% higher pricing on older-generation GPUs than in Q1.
• Four landmark deals averaging more than $1B of TCV each, including Reflection and Cohere.
• TCV closed in Q2 grew nearly 4x sequentially; TCV from new customers grew more than 9x.
• In early Q3, a first short-term capacity deal and a first capacity auction pilot, with a stated price opportunity of $40–50M per MW.

Prices for AI compute are going up, on old hardware as well as new. That is the assumption the entire neocloud sector's capex is underwritten by, and this is one of the cleaner disclosures of it.

The guidance is not in the filing

Contracted power guidance was raised again, to 5 GW — the fifth raise in twelve months (>1 GW in Aug '25, >2.5 GW, >3 GW, >4 GW, now 5 GW). Full-year 2026 guidance was "reiterated across all metrics," with the actual numbers left to the call. They are worth stating, because no SEC filing contains them:

Put the first and fourth rows together: Nebius has guided to spending roughly six to eight times its full-year revenue on capex in 2026. And the ARR line implies the $3.0B run-rate reached in June roughly triples again by December.

There is a modelling trap in the power figure that the call corrected. The 800 MW – 1 GW is connected power, described as active through the first half of 2027 — not revenue-producing capacity on December 31. Almost all of the 5 GW contracted target arrives over the next two to three years.

The most revealing sentence on the call

Volozh, explaining why Nebius is not selling its 2027 capacity:

"Most importantly, we could sell today our entire 2027 capacity on these terms if we wanted to. We are not doing this."

The company is deliberately withholding inventory in a shortage, holding back premium capacity for short-duration deals at $40–50M per MW against the $20–25M on the four landmark Q2 deals. Management described prioritising terms in the order price, then prepayment, then duration, and framed the platform as choosing "when to sell, to whom we sell, and on what terms."

The supporting datapoint is the sharpest pricing evidence in the sector this quarter: a capacity auction cleared 15% above the highest previous Blackwell price Nebius had achieved. Both the auction and the short-term deals come online later in 2026, so management was explicit that they do not affect 2026 revenue guidance — they land in 2027.

A seller that could clear its entire forward book today and chooses not to is making a stronger statement about price direction than any forecast.

What to watch

Shares rose about 16.5% on the print.

Three things decide whether the next few quarters look like this one:

1. Does ARR keep converting? $3.0B of run-rate against $582.3M of quarterly revenue implies the recognized number roughly doubles from here on capacity already contracted, and the guide asks for $7–9B of ARR by December. Management says most Q2 deals were signed against capacity arriving in late 2026 and will contribute primarily to 2027 revenue — so the conversion is scheduled, not immediate. H1 revenue was $981.3M against a $3.0–3.4B full-year guide, which requires the second half to roughly double the first.
2. Does the prepayment share hold? 70% is an all-time high, and it is what makes a −$3.4B free-cash-flow quarter financeable. If it reverts, the funding mix shifts to debt and equity at exactly the moment capex is largest.
3. Does the pricing hold? The bull case is now explicitly priced at $20–25M per MW with $40–50M available on short-dated capacity. Every element of the payback math depends on that not being the top.

The comparison set reports around the same shape: CoreWeave's Q2 showed the same collision of positive operating cash flow, enormous capex and interest expense eating the operating line. Nebius is the smaller company with the faster growth rate and — for now — the better margin.

What we learned

1. The comparison base changed, and the growth rate depends on which one you use. After the Toloka deconsolidation, prior periods were reclassified to discontinued operations — 2025 Q2 went from $152.8M to $105.1M. The 454% headline is right on the restated base; against the old figure it is 281%.
2. Operating leverage is genuinely arriving. Every expense line fell as a share of revenue — cost of revenues 29% → 23%, product development 41% → 33%, SG&A 65% → 30%, D&A 72% → 45% — taking total operating costs from 206% of revenue to 130%.
3. The cloud earns more than the company does. Nebius AI cloud revenue of $574.9M, about 98% of the total, carried $285.7M of segment adjusted EBITDA at a 49.7% margin — more than the consolidated $236.2M, meaning Avride and TripleTen cost roughly $50M of EBITDA in the quarter.
4. $5.66 billion of property and equipment in three months, against $582.3M of revenue. Nearly ten times the quarter's entire top line.
5. The capital structure is now visible in the P&L. Interest expense went from $4.8M to $119.1M and D&A from $75.2M to $259.7M, turning a 41% adjusted EBITDA margin into a $190.4M net loss.